In [1]:
###################################################################
# Built-in modules ################################################
###################################################################
import sys
sys.path.append("../../")
import os
import numpy as np
import imp
import cv2
import os
import matplotlib.pyplot as plt
import gc
import math
import time
import glob
###################################################################
# Custom modules ##################################################
###################################################################
from components.matchers.NumbaPatchMatcher import Wrapper as m
from components.utils.SimpleProgressBar import SimpleProgressBar
from components.utils import utils as u
from components.utils import plot_utils as plu
from components.utils.CSVWriter2 import Wrapper as csv
from components.utils.Metrix import Wrapper as me
###################################################################
# Dataset specific modules#########################################
###################################################################
from components.utils import middlebury_utils as mbu
In [2]:
ROOT_PATH = os.path.join("..", "..")
EXPERIMENT_TITLE = "ALG_005_EXP_002-PatchMatch-MacLean_et_al-Numba"
INIT_METHOD = "maclean"
DATASET = "middlebury"
DATASET_FOLDER = os.path.join(ROOT_PATH, "datasets", DATASET)
LOG_FOLDER = os.path.join(ROOT_PATH, "experiments", "logs")
CSV_FILEPATH = os.path.join(LOG_FOLDER, EXPERIMENT_TITLE+".csv")
IMG_RES = "450X375"
PREPROCESSING_METHOD = "None"
KERNEL_SIZE = 1
KERNEL_SPEC = "None"
SCENES = ["teddy", "cones"]
SIZE=""
YEAR= 2003
EXP_PARAMS = dict()
Loading the dataset and visualising the results¶
In [3]:
loaded_imgs_and_paths = list(mbu.get_images(DATASET_FOLDER, YEAR, scene) for scene in SCENES)
for im, path in loaded_imgs_and_paths:
filenames = list(os.path.split(p)[-1] for p in path)
plu.plot_images(im, filenames)
The addition of occluded areas to the pristine groundtruth images¶
In [4]:
###################################################################
# 2. Occluded areas are added to the groundtruths #################
###################################################################
nonocc_gts = list()
for imgs, paths in loaded_imgs_and_paths:
temp = mbu.add_masks_to_raw_disparity(imgs[2], imgs[3])
nonocc_gts.append(mbu.add_masks_to_raw_disparity(imgs[2], imgs[3]))
plu.plot_images(nonocc_gts, ["teddy_nonocc_gt", "cones_nonocc_gt"])
Benchmarking and evaluation¶
In [5]:
###################################################################
# 3.1 Initialising hyperparameters and matcher#####################
###################################################################
csv_logger = csv(CSV_FILEPATH, default_header=False)
csv_logger.set_header_function(csv_logger.get_header_v3)
csv_logger.write_csv_header()
csv_logger.set_line_function(csv.format_stereo_matching_results_v2)
MATCH = 60
GAP = -20
EGAP = -1
EXP_PARAMS = {"experiment_id":EXPERIMENT_TITLE, "match": MATCH,"gap":GAP,"egap":EGAP,\
"algo":str(m.__module__), "init_method":"default","dataset":DATASET,\
"preprocessing_method":"None","kernel_size":1,"kernel_spec":"None"}
In [6]:
###################################################################
# 3.2 Matching ####################################################
###################################################################
from skimage.morphology import disk, diamond, star, octagon
from skimage import filters
results = []
EXP_PARAMS["init_method"] = "maclean_et_al"
patches = dict(
#Gaussian filters
gauss_3_3 = u.gaussian_kernel(3,3,1,1)*9,
gauss_5_5 = u.gaussian_kernel(5,5,1,1) *25,
gauss_5_3 = u.gaussian_kernel(5,3,1,1) *15,
gauss_7_3 = u.gaussian_kernel(7,3,1,1) *21,
gauss_11_3 = u.gaussian_kernel(11,3,1,1) *33,
gauss_13_3 = u.gaussian_kernel(13,3,1,1) *39,
#Sci-image morphological
star_2=star(2),
star_3=star(3),
octagon_1_3 = octagon(1,3),
octagon_3_1 = octagon(3,1),
diamond_2 = diamond(2),
diamond_3=diamond(3),
disk_2=disk(2),
disk_3=disk(3),
disk_4=disk(4),
#Farid filters
hfarid = filters.edges.HFARID_WEIGHTS,
vfarid = filters.edges.VFARID_WEIGHTS,
hvfarid_div2 = (filters.edges.HFARID_WEIGHTS+filters.edges.VFARID_WEIGHTS)/2,
hvfarid = (filters.edges.HFARID_WEIGHTS+filters.edges.VFARID_WEIGHTS),
#Sobel filters
hsobel = filters.edges.HSOBEL_WEIGHTS,
vsobel = filters.edges.VSOBEL_WEIGHTS,
hvsobel_div2 = (filters.edges.HSOBEL_WEIGHTS+filters.edges.VSOBEL_WEIGHTS)/2,
hvsobel_div = (filters.edges.HSOBEL_WEIGHTS+filters.edges.VSOBEL_WEIGHTS),
#Scharr filters
hscharr = filters.edges.HSCHARR_WEIGHTS,
vscharr = filters.edges.VSCHARR_WEIGHTS,
hvscharr_div2 = (filters.edges.HSCHARR_WEIGHTS+filters.edges.VSCHARR_WEIGHTS)/2,
hvscharr = (filters.edges.HSCHARR_WEIGHTS+filters.edges.VSCHARR_WEIGHTS),
#Prewitt filters
hprewitt = filters.edges.VPREWITT_WEIGHTS,
vprewitt = filters.edges.VPREWITT_WEIGHTS,
hvprewitt_div2 = (filters.edges.VPREWITT_WEIGHTS+filters.edges.HPREWITT_WEIGHTS)/2,
hvprewitt = (filters.edges.VPREWITT_WEIGHTS+filters.edges.HPREWITT_WEIGHTS)
)
In [9]:
matcher = m(MATCH, GAP, EGAP)
progress_bar = SimpleProgressBar.get_instance()
progress_counter = 1
steps_to_be_taken = len(SCENES)*len(patches)*4
print("Number of iterations: {0}".format(steps_to_be_taken))
SAVE_PATH = os.path.join(ROOT_PATH, "experiments", "disparities", EXPERIMENT_TITLE)
saved_image_names = []
results = []
In [10]:
for match_value in range(10, 41, 10):
for k,p in patches.items():
for i, scene in enumerate(SCENES):
# Parameters to be logged...
EXP_PARAMS["scene"] = scene
EXP_PARAMS["image_filename"] = IMAGE_FILENAME = loaded_imgs_and_paths[i][1][0]
EXP_PARAMS["img_res"] = "{0}x{1}".format(loaded_imgs_and_paths[i][0][1].shape[1], loaded_imgs_and_paths[i][0][1].shape[0])
EXP_PARAMS["kernel_size"] = "{0}x{1}".format(p.shape[0], p.shape[1])
EXP_PARAMS["kernel_spec"] = k
EXP_PARAMS["match"] = MATCH = match_value
matcher = m(MATCH, GAP, EGAP)
occ = loaded_imgs_and_paths[i][0][3]
gt = loaded_imgs_and_paths[i][0][2]
matcher.set_images(loaded_imgs_and_paths[i][0][0].astype(np.float64),
loaded_imgs_and_paths[i][0][1].astype(np.float64))
tic = time.time()
matcher.configure_instance()
matcher.set_filter(p)
tic=time.time()
x,raw_disp_map = matcher.test_pipeline()
toc = time.time()
EXP_PARAMS["runtime"] = toc-tic
disp = raw_disp_map
disp = 4*disp
temp_path = u.save_disparity(SAVE_PATH, disp)
saved_image_names.append(os.path.split(temp_path)[-1])
EXP_PARAMS["image_filename"] = temp_path
EXP_PARAMS["are_occlusions_errors"] = ARE_OCCLUSIONS_ERRORS = False
EXP_PARAMS["bad1"], EXP_PARAMS["bad2"], EXP_PARAMS["bad4"], EXP_PARAMS["BAD8"], EXP_PARAMS["abs_error"], EXP_PARAMS["mse"], EXP_PARAMS["avg"], EXP_PARAMS["eucledian"] =\
BAD1, BAD2, BAD4, BAD8, ABS_ERR, MSE, AVG, EUCLEDIAN = me.evaluate_over_all(disp, gt, occ, occlusions_counted_in_errors = ARE_OCCLUSIONS_ERRORS)
csv_logger.append_new_sm_results(EXP_PARAMS, selected_keys = csv.get_header_v3())
EXP_PARAMS["are_occlusions_errors"] = True
EXP_PARAMS["bad1"], EXP_PARAMS["bad2"], EXP_PARAMS["bad4"], EXP_PARAMS["BAD8"], EXP_PARAMS["abs_error"], EXP_PARAMS["mse"], EXP_PARAMS["avg"], EXP_PARAMS["eucledian"] =\
BAD1, BAD2, BAD4, BAD8, ABS_ERR, MSE, AVG, EUCLEDIAN = me.evaluate_over_all(disp, gt, gt, occlusions_counted_in_errors = ARE_OCCLUSIONS_ERRORS)
csv_logger.append_new_sm_results(EXP_PARAMS, selected_keys = csv.get_header_v3())
progress_bar.progress_bar(progress_counter, steps_to_be_taken, header= "Experiment on patch sizes in progress: ", progress_bar_steps=40)
progress_counter+=1
results.append(disp)
In [11]:
import pandas as pd
experiment_log = pd.read_csv(CSV_FILEPATH)
experiment_log.sort_values(by="abs_error", inplace=True)
experiment_log.head(30)
Out[11]:
In [12]:
plu.plot_images(results, saved_image_names, ncols=2)
